Cardinality estimation, which involves estimating the result size of queries, is a critical aspect of query processing and optimization. Deep Neural Networks (DNNs) are data hungry, and being trained directly for cardinality estimation of activity trajectory similarity queries usually leads to poor performance. To address this problem, we propose two enhancements to improve accuracy and reduce the training data size: query slice and data slice. Query slice divides a query trajectory into three slices based on similarity dimensions (i.e., spatial, temporal, and textual). Data slice organizes similar trajectories into data slices. Furthermore, we design a global-local model to determine which local models should be used for a given query. We also extend above model to support similarity joins. Experimental results show that our method efficiently learns to estimate the cardinality, achieving superior accuracy and efficiency compared to state-of-the-art methods.
Read full abstract